NVlabs/RoboLab

A simulation benchmarking platform for evaluating generalist robot policies

What it solves

RoboLab provides a standardized, high-fidelity simulation environment for evaluating generalist robot manipulation policies. It addresses the need for reproducible, large-scale benchmarking by offering a vast library of tasks with automated success detection, removing the manual effort required to verify if a robot has successfully completed a complex manipulation goal.

How it works

Built on NVIDIA Isaac Lab, RoboLab implements a server-client architecture where the robot policy (the AI model) runs as a standalone server and connects to the simulation via a lightweight inference client. This setup allows for multi-environment parallel evaluation, meaning many episodes can be run simultaneously to speed up testing. The system uses composable predicates for automated success/failure detection and includes a web-based dashboard for replaying episode videos and analyzing cross-experiment results.

Who it’s for

It is designed for researchers and developers building generalist robot policies who need a rigorous way to test their models across a wide variety of manipulation tasks without being tied to a specific robot hardware embodiment.

Highlights

  • RoboLab-120: A benchmark of 120 tasks including pick-and-place, stacking, and tool use, all with language instructions.
  • Robot Agnostic: Compatible with any robot that works with IsaacLab.
  • AI-Assisted Creation: Includes Claude Code skills to generate new scenes and tasks using natural language.
  • Parallel Evaluation: Supports vectorized conditionals and per-environment termination for high-throughput testing.
  • Integrated Dashboard: A dedicated web interface for browsing tasks and analyzing experiment results.

Related

  • Project
    georghess/neurad-studio

    neurad‑studio is a research‑grade library that brings neural‑rendering (NeRF and 3‑D Gaussian splatting) to autonomous‑driving data. It supplies ready‑to‑run implementations of NeuRAD and SplatAD, parsers for major AD datasets, a viewer that can display both camera images and lidar points, and standard training/visualisation tools built on Nerfstudio.

  • Project
    umfieldrobotics/OceanSim

    A GPU-accelerated underwater robot perception simulation framework that uses physics-based rendering to model visual and acoustic sensors and reduce the simulation-to-real gap.

  • Project
    NVIDIA-ISAAC-ROS/isaac_ros_pose_estimation

    A collection of NVIDIA-accelerated ROS 2 packages for 3D object pose estimation, enabling robots to locate and track both known and novel objects for navigation and manipulation.

  • Project
    ros2/rmw_zenoh

    A ROS 2 RMW implementation based on Zenoh that provides an alternative to DDS for robot communication, featuring implicit shared memory and flexible cross-host discovery.